Know how good your AI agents’ decisions are, before they count.

No. The operating agent makes decisions using its own instructions, models, tools and context. The Blueprint independently represents how your organization expects consequential decisions to be made. Steward compares the agent’s observable decision trajectory against that organizational reference. Nor does the agent need to imitate any one person: different people reach sound decisions by different paths. The Blueprint holds what any sound path must include, and Steward compares only what is observable or evidenced.

No. The workflow tells us what the agent does. The decision trajectory shows how it arrives at a decision: what is considered, which conditions matter, what is omitted and when escalation is needed. The Blueprint makes visible the judgment the organization expects. Most companies have the first, not the second. The sprint keeps one agent workflow as its scope and builds the second within it, starting from your policies, playbooks and experts.

Each of them checks something else. Risk checks policy and exposure after the fact; QA checks that the agent does what the spec says; AI engineering checks the model against benchmarks. Whether a decision weighed what your best people would have weighed sits in none of those remits, and the person who used to answer it by instinct is the one the agent replaced. Your teams could build this. Nobody has asked them to, and they would do it after the first incident.

Those describe the system: which agents you run, who built them and who is accountable for them. The Blueprint describes its decisions: each one the agent takes, who may stop it, who answers for it, and the criteria it must meet, agreed by every team it touches. A quick test: take a decision your agent made last Tuesday. Who answered for it, and against which criteria? If your documents answer that, you have it already. See the comparison →

Yes, and the self-assessment shows where to start. The hard part is not the document but the agreement: getting the teams involved to surface where they contradict each other, and sign. That stalls for months internally, because nobody owns the space between teams. The sprint brings the method, an outside facilitator and a three-week deadline.

Your evaluation and policy tools can enforce business criteria, once someone has written them down. What they lack is the standard itself: the criteria for each decision, owned by the business and agreed by every team involved. That is what the Blueprint establishes. Your existing tools can then enforce it, and Steward checks each decision against it as the agent runs.

A decision gap assessment of the decision trajectories in one agent workflow: how the agent decides today, what your existing controls cover, and what they miss. Then one shared, endorsed Agent Decision Blueprint: conflicting criteria resolved, a named owner for each decision, and the points where people step in. A replay of past agent decisions shows which would have been flagged or sent to an owner. You also keep a standard your team can reuse on the next workflow, and a report for your board.

No. Steward runs as a wrapper, sidecar or gateway around the agent. It watches first; nothing is enforced until the owners switch it on.

Your people. The Blueprint is filled by the people who held that judgment before the agent did, and each check is endorsed by a named owner. We bring the method and the instrument; the standard is yours.

Steward does not make the business decision; it executes the response the owner endorsed for that case: pass, nudge, throttle, flag, contain, escalate or block. When the endorsed response is to escalate, the decision goes to the owner of the check it failed, with what was weighed and what was missing. The owner decides, and the Blueprint grows if they learn something.

An LLM judge rates an output against a rubric it holds, and its verdict is the end of the story. Steward measures the distance between what the agent weighed and a decision basis your people wrote, endorsed and version-controlled; it does not hold an opinion of its own. What follows a miss is not the model’s call either: it is the response the owner endorsed for that check. And every call Steward makes is in the record, so its own error rate is measured, not trusted. The model is the instrument. The judgment is yours.

We test this directly. In one control, a Blueprint required triage agents to screen patients for alien abduction. The agent worked competently and never asked. Steward flagged exactly that gap: it applies your standard, however unusual, and nothing else.

Any system can run a check. The question is whether you know when the check is wrong.

Steward applies your people’s judgment at machine speed. It does not define what a good decision is; it checks each agent decision against the business criteria your people endorsed, records what it finds, and makes the gaps visible. Before it is allowed to intervene, we test it against decisions reviewed by your own experts: what it catches, what it misses, where it raises a false alarm.

Because they are tested, not assumed. In week three, a sample of the agent’s past decisions is replayed against the Blueprint and your experts review what would have been escalated or routed. Criteria that flag sound decisions get tuned; considerations that were missing get added. Once Steward runs, the record shows when a criterion stops matching reality.

As a gateway, Steward can hold or block any decision that passes through it. As a wrapper or sidecar it flags, escalates and contains, which for most workflows is what the owner wants first. You choose how far it goes, per workflow.

Where you choose. You pick the model that verifies and where Steward runs; the decision basis and the record stay with you.

One decision the agent makes often, that costs something when it goes wrong, and that has an owner who cares. Claims, supplier qualification and release approval are the three on this page; on the call we find yours.

No. The claims case leads because it is the easiest to follow in a minute. The same kind of Blueprint works for procurement, releases, pricing, credit, support refunds: any workflow where an agent now makes a decision a person used to make.

A fixed fee for the sprint, agreed on the call once we know the workflow. If after week one you do not see the value, you stop and pay for week one only. After the sprint, Steward continues per workflow.

A scoped test, not a platform migration: one workflow, three weeks, no agent rewrite, and a week-one exit. You leave knowing whether the gap is real in your organization. Either answer is worth having.

The replay runs on decision records you select and export. No system access, and nothing touches production. Your data is used only for the sprint and deleted when it ends. MeaningStack is a Dutch B.V.; processing is covered by our DPA, and we are happy to work on your paper.

It checks every decision against the considerations your people named, and flags a decision that rests on something nobody put on the Blueprint. It also catches an agent that is not what it claims to be: a swapped or compromised agent may pass authentication, but its decisions stop matching your criteria.

It cannot tell you whether the unknown was right. That judgment stays with your people, and once they decide, the Blueprint grows to include it.